English

Image Classification at Supercomputer Scale

Machine Learning 2018-12-04 v2 Distributed, Parallel, and Cluster Computing Machine Learning

Abstract

Deep learning is extremely computationally intensive, and hardware vendors have responded by building faster accelerators in large clusters. Training deep learning models at petaFLOPS scale requires overcoming both algorithmic and systems software challenges. In this paper, we discuss three systems-related optimizations: (1) distributed batch normalization to control per-replica batch sizes, (2) input pipeline optimizations to sustain model throughput, and (3) 2-D torus all-reduce to speed up gradient summation. We combine these optimizations to train ResNet-50 on ImageNet to 76.3% accuracy in 2.2 minutes on a 1024-chip TPU v3 Pod with a training throughput of over 1.05 million images/second and no accuracy drop.

Keywords

Cite

@article{arxiv.1811.06992,
  title  = {Image Classification at Supercomputer Scale},
  author = {Chris Ying and Sameer Kumar and Dehao Chen and Tao Wang and Youlong Cheng},
  journal= {arXiv preprint arXiv:1811.06992},
  year   = {2018}
}

Comments

Presented as part of Systems for ML Workshop @ NIPS 2018

R2 v1 2026-06-23T05:18:37.215Z